ML Documentation: what's missing

[kingjr]

Here is a proposal in the todo for the MVPA/Decoding-Encoding/ML module.

• a brief tutorial on ML to introduce the key concepts (fitting, predicting, scoring, CV)
• a tutorial on model interpretation and its limits: when can you do patterns, what does it mean, what are the risks
• a tutorial to highlight the importance of supervised spatial filter for oscillatory activity
• a tutorial to highlight importance of supervised spatial filters for various types of artefacts
• a tutorial to do source decoding and its risks
• a tutorial to explain pipeline with 2D, 3D, 4D tensors

Please comment and I'll update the proposal.

[agramfort]

do you have already materials for some of this?

[jasmainak]

It's more an API thing but I get a lot of questions which basically boil down to using 3D data in pipelines leading to weird shape mismatches. I guess some estimators can be pipelined in sequence but others cannot be. This needs to be probably clarified.

[kingjr]

do you have already materials for some of this?

Yes, in the form of ipynb for all of the above, but a bit too long as compared to the MNE tutorials. Shall I start opening PRs?

It's more an API thing but I get a lot of questions which basically boil down to using 3D data in pipelines leading to weird shape mismatches. I guess some estimators can be pipelined in sequence but others cannot be. This needs to be probably clarified.

Ok that's easy, we can dedicate a tutorial to pipeline. I think it is an important an concept, especially with our tensor-like data structure.

[agramfort]

let's not make PRs but maybe make a repo of 3-4 notebooks with an index.ipynb like here:

https://github.com/jakevdp/PythonDataScienceHandbook/tree/master/notebooks

then we take the figures from the notebook for the paper.